Skip to main content

Osaurus

Osaurus

Own your AI.

Install OsaurusQuick Start

Where do you want to go?​

Get started

Install in under a minute, create your first agent, and pick a model — local, Apple Foundation, or cloud.

Use Osaurus

Chat from anywhere with ⌘;, build agents for different jobs, add memory, skills, voice, and image generation.

Build with Osaurus

OpenAI-, Anthropic-, and Ollama-compatible local server, MCP in and out, a CLI, and a stable plugin ABI.

Privacy & trust

Everything stays on your Mac, signed at every boundary. See exactly what is stored and what leaves — and when.

Inference is all you need. Everything else can be owned by you.​

Models are getting cheaper and more interchangeable by the day. What's irreplaceable is the layer around them — your context, your memory, your tools, your identity. Other apps keep that layer on their servers. Osaurus keeps it on your Mac.

Osaurus is the AI harness for macOS. It sits between you and any model — local or cloud — and provides the continuity that makes AI personal: agents that remember, execute autonomously, run real code, and stay reachable from anywhere.

Works fully offline with local models. Connect to any cloud provider when you want more power. Nothing leaves your Mac unless you choose. Native Swift on Apple Silicon. No Electron. MIT licensed.

Your data, your Mac

Stored only on your Mac, signed at every boundary, never sent anywhere unless you explicitly choose a cloud provider. We can't read your conversations and there are no backdoors — see Security & Privacy.

What you get​

A short tour of the things you can do once Osaurus is installed.

Everyday AI​

  • A guided first launch. A short onboarding walks you through naming your first agent and giving it a brain — the best local model for your Mac is picked automatically, with Osaurus Cloud included via a free welcome credit, or bring your own API key. Identity is set up silently. No config files.
  • A chat overlay you can open anywhere. Press ⌘; to talk to your AI; press it again to dismiss. No browser tab, no context switch.
  • Agents that fit different jobs. A coding partner, a research assistant, a file organizer — each with its own prompt, theme, and history.
  • Voice input. Dictate in chat, talk to an agent hands-free with a wake word, or hold a hotkey to dictate into any app — all on-device.
  • Image generation, fully offline. Install a local image model and create or edit pictures right in the chat — nothing is sent to a server.
  • Themes. Built-in light/dark, fully editable, importable as JSON.

Memory & knowledge​

  • Memory that learns from you. Past conversations are distilled into compact facts and surfaced only when relevant — no firehose of irrelevant context.
  • Skills that load themselves. Packaged expertise — research, debugging, writing styles — that surface automatically when the task calls for them.
  • Web search, no setup. Every agent can search the web out of the box — keyless built-in providers, optional API keys for better quality. See Web Search.
  • Working folders. Point a chat at a folder and the agent gets safe file, search, and git tools — scoped to just that directory.

Autonomy & automation​

  • A Sandbox for real code. Toggle it on and the agent can run shell, Python, and Node in isolation — a Linux VM on macOS 26+, a Seatbelt-confined runner on macOS 15.
  • Computer Use (experimental). Let an agent drive real macOS apps — fill forms, flip settings, extract on-screen text — with every action gated by a safe-by-default confirmation policy.
  • Browser Use. Give an agent its own persistent, isolated browser — it navigates, reads, and fills forms, with the same safe-by-default action gating. See Browser Use.
  • Subagents. Delegate a bounded task to another model or a saved agent mid-conversation and get a compact result back — local models hand off memory automatically.
  • Schedules and Watchers. Run an agent on a timer, or whenever a folder changes. Useful for daily journals, screenshot organizers, end-of-day commits.

Privacy & reach​

  • A Privacy Filter for cloud prompts (experimental). An on-device classifier scrubs names, emails, and secrets out of anything you send to a cloud provider — you review what's redacted, and it blocks the send rather than leak.
  • Identity that's yours. A cryptographic address for you and each of your agents. Issue access keys for outside tools, scope them per-agent, revoke them whenever.
  • Public links without ports. Give one agent a stable public URL via a secure tunnel through agent.osaurus.ai — no port forwarding, no ngrok.

Get started in 5 minutes →

Build with Osaurus​

Osaurus is also a local server. It speaks OpenAI, Anthropic, Open Responses, and Ollama APIs at the same port — so any SDK you already use just works. And it's a full MCP server and client, so Cursor, Claude Desktop, and other MCP harnesses get instant access to your installed tools.

  • HTTP API — endpoint reference, streaming, function calling
  • SDK Examples — Python, JavaScript, Anthropic SDK, Open Responses
  • CLI — osaurus serve, osaurus tools install/dev/create, osaurus mcp
  • Tools & Plugins — 20+ native plugins (Mail, Calendar, Vision, Git, …) and a stable v1–v6 ABI for building your own
  • Apple Intelligence — using foundation with zero setup on macOS 26+

For the system view of how everything fits together, see Architecture.

System requirements​

  • macOS 15.5 or later
  • Apple Silicon (M1, M2, M3, or newer)
macOS 26 features

Apple Foundation Models require macOS 26 (Tahoe) or later, and the Sandbox uses its full Linux VM there — on macOS 15 the Sandbox falls back to a Seatbelt-confined backend.

Community​

Osaurus is an indie project, built in public. Join us:

  • Discord — chat, feedback, show-and-tell
  • GitHub — issues, contributions, roadmap
  • Hugging Face — curated quantizations for Apple Silicon
  • Plugin Registry — browse and submit tools
  • Blog — long-form thinking on personal AI